NBER WORKING PAPER SERIES
WHY DO MOTHERS BREASTFEED GIRLS LESS THAN BOYS? EVIDENCE AND
IMPLICATIONS FOR CHILD HEALTH IN INDIA
Seema Jayachandran
Ilyana Kuziemko
Working Paper 15041
http://www.nber.org/papers/w15041
NATIONAL BUREAU OF ECONOMIC RESEARCH
1050 Massachusetts Avenue
Cambridge, MA 02138
June 2009
We thank Kristin Plys and Luke Stein for excellent research assistance and Anne Case, Monica Das
Gupta, Angus Deaton, Pascaline Dupas, Claudia Goldin, Jeffrey Hammer, Larry Katz, Alan Krueger,
Adriana Lleras-Muney, Ben Olken, Emily Oster, Eliot Regett, Shanna Rose, Sam Schulhofer-Wohl,
Chris Woodruff, and conference and seminar participants at Princeton, Stanford, the Pacific Conference
for Development Economics, and the NBER Children's meeting for helpful comments. The views
expressed herein are those of the author(s) and do not necessarily reflect the views of the National
Bureau of Economic Research.
NBER working papers are circulated for discussion and comment purposes. They have not been
peer-reviewed or been subject to the review by the NBER Board of Directors that accompanies official
NBER publications.
© 2009 by Seema Jayachandran and Ilyana Kuziemko. All rights reserved. Short sections of text, not
to exceed two paragraphs, may be quoted without explicit permission provided that full credit, including
© notice, is given to the source.
Why Do Mothers Breastfeed Girls Less Than Boys? Evidence and Implications for Child
Health in India
Seema Jayachandran and Ilyana Kuziemko
NBER Working Paper No. 15041
June 2009
JEL No. I1,J13,O12,O15
ABSTRACT
Medical research indicates that breastfeeding suppresses post-natal fertility. We model the implications
for breastfeeding decisions and test the model's predictions using survey data from India. First, we
find that breastfeeding increases with birth order, since mothers near or beyond their desired total fertility
are more likely to make use of the contraceptive properties of nursing. Second, given a preference
for having sons, mothers with no or few sons want to conceive again and thus limit their breastfeeding.
We indeed find that daughters are weaned sooner than sons, and, moreover, for both sons and daughters,
having few or no older brothers results in earlier weaning. Third, these gender effects peak as mothers
approach their target family size, when their decision about future childbearing (and therefore breastfeeding)
is highly marginal and most sensitive to considerations such as ideal sex composition.
Because breastfeeding protects against water- and food-borne disease, our model also makes predictions
regarding health outcomes. We find that child-mortality patterns mirror those of breastfeeding with
respect to gender and its interactions with birth order and ideal family size. Our results suggest that
the gender gap in breastfeeding explains 14 percent of excess female child mortality in India, or about
22,000 "missing girls" each year.
Seema Jayachandran
Department of Economics
Stanford University
579 Serra Mall
Stanford, CA 94305-6072
and NBER
jayachan@stanford.edu
Ilyana Kuziemko
Princeton University
361 Wallace Hall
Princeton, NJ 08544
and NBER
kuziemko@princeton.edu
1
Introduction
As medical and public health researchers have long documented, breastfeeding inhibits post-natal fertility. The converse also holds: the physical demands of another pregnancy often cause a mother to stop nursing her current child. These biological constraints suggest a negative relation-ship between breastfeeding duration and future fertility.1
In developing countries, this relationship may be particularly important. First, many women lack access to modern forms of birth control, and thus rely more heavily on the contraceptive properties of breastfeeding. Second, mothers have more difficulty meeting the high caloric demands of breastfeeding while pregnant in environments characterized by high rates of malnourishment. Finally, breastfeeding is thought to have greater health benefits for infants who would otherwise consume unsafe drinking water and contaminated food.2
This paper studies the relationship between breastfeeding and future fertility in India. We test whether factors that likely affect a mother’s decision to have more children—such as her current number of children and their sex composition—affect the length of time she nurses. We also examine the health consequences for children who are breastfed less, e.g., low-birth-order children and girls.
Of course, many factors aside from subsequent fertility could affect a mother’s propensity to breastfeed, such as her health, labor-market attachment, and education, or the price and availabil-ity of breast-milk substitutes. To help separate our hypothesis from these alternatives, we specify a dynamic programming model of breastfeeding as a function of desired future fertility. The model makes a number of distinct predictions about breastfeeding patterns with respect to a child’s gen-der, siblings’ sex composition, and birth order. We then test these predictions using data from the 1992, 1998 and 2005 waves of the National Family Health Survey in India. While we do not try to eliminate every alternative hypothesis, our model predicts (and the data support) a number of very specific empirical patterns that any alternative hypothesis would also have to explain.
First, breastfeeding increases with birth order. As mothers reach their ideal family size, their desire for more children falls and thus their demand for breastfeeding rises. Second, so long as some preference for sons exists, boys are breastfed more than girls; after the birth of a daughter, mothers are more likely to continue having children (and thus limit breastfeeding) in order to try for a son. Third, by the same logic, children with older brothers are breastfed more. Fourth, 1Research suggests that nursing suppresses the Gonadotropin-releasing hormone (GnRH) that regulates ovulation. Nursing can also cause weight loss, which can disrupt ovulation (Blackburn, 2007). While most doctors believe that there is no danger in breastfeeding while pregnant, many women in our data cite pregnancy as the reason they stopped nursing. See the next section for further discussion.
2
these gender effects are smallest for high and low values of birth order. For low (high) birth-order children, mothers will want to continue (stop) having children regardless of the sex of her children and thus breastfeed boys and girls equally. Finally, the gender effect should be largest when a couple approaches its target family size, as at that point their decision to have another child is highly marginal and thus most sensitive to sex composition.
Our results have several potentially important implications for researchers and policymakers. First, breastfeeding’s contraceptive effects, coupled with its potential health benefits to children, suggest a negative relationship between total fertility and child health. That is, a mother who wants to have many children will wean her child sooner in order to conceive again quickly, with negative consequences for the child’s health. Conversely, a mother who uses breastfeeding to limit her family size will confer health benefits to her children. Thus, breastfeeding represents a heretofore unexamined mechanism for the quantity-quality tradeoff in fertility introduced by Becker (1960), Becker and Lewis (1973) and Becker and Tomes (1976), and documented in India by Rosenzweig and Wolpin (1980), among others. Through this mechanism, as fertility rates fall, mothers will, on average, breastfeed their children more.
Second, the combination of the health benefits of breastfeeding and the gender gap in breast-feeding may help explain the “missing girls” problem in India (Sen, 1990). Boys are breastfed more than girls, but the difference is small for first-borns and in the first six months of life, which coin-cides with excess female mortality patterns in India. We find that excess female mortality between ages one and three years—the age range where there is a large gender gap in breastfeeding—is driven by households without piped water, which is consistent with girls’ lower rate of breastfeeding increasing their exposure to water-borne disease and, in turn, their mortality rate. Back-of-the-envelope calculations suggest that breastfeeding accounts for 14 percent of the gender gap in child mortality (deaths between ages one and five) in India, or 22,000 missing girls each year. Son pref-erence is the underlying cause of this excess female mortality, but in a subtle way: Rather than resulting from parents’ explicit decisions to allocate more resources to sons, the missing girls are mainly an unintended consequence of parents’ desire to have more future sons.
Third, our results have potential policy implications related to modern contraception. Expand-ing the availability of birth control could either increase or decrease breastfeedExpand-ing. If mothers rely on breastfeeding when more effective forms of contraception are unavailable, then access to mod-ern birth control might lead them to substitute away from breastfeeding. Although the benefits of modern contraception may well swamp this potential cost, its introduction may need to be coupled with campaigns to encourage breastfeeding if policy makers wish to prevent a decline in breastfeeding rates. Moreover, improving water quality may become a more urgent policy priority
in communities with access to contraception, given the possible declines in nursing. Conversely, modern contraception could increase breastfeeding if, by giving mothers greater control over the timing and number of pregnancies, it reduces their need to suspend breastfeeding due to a new pregnancy.
The remainder of the paper is organized as follows. Section 2 provides background on the rela-tionship between breastfeeding and fertility. Section 3 presents the model, and Section 4 describes the data. Section 5 tests the model’s predictions that breastfeeding increases with birth order. Section 6 tests the predictions that breastfeeding depends on the sex composition of a mother’s children and that this effect interacts with birth order. Section 7 discusses the health effects re-lated to these breastfeeding patterns. Section 8 discusses how access to modern contraception might affect breastfeeding and offers concluding remarks.
2
Background on breastfeeding and subsequent fertility
2.1 Are women less fertile while they breastfeed?
The medical research suggests at least two mechanisms by which breastfeeding inhibits fertility. First, nursing affects certain hormones that regulate ovulation. Breastfeeding appears to interrupt the release of the Gonadotropin-releasing hormone (GnRH), which triggers the pituitary gland to release high levels of luteinizing hormone (LH). This so-called “LH surge” marks the beginning of ovulation. There is also some evidence that breastfeeding increases levels of the hormone prolactin, which inhibits ovulation (Blackburn, 2007).
Second, nursing diverts calories from the mother to the infant. For mothers who consume a limited number of calories, this diversion can lead to malnutrition, which shuts down ovulation. This channel likely plays an important role in developing countries. Indeed, using India’s National Family Health Survey (NFHS), we regress mother’s weight on a dummy variable for whether she is currently breastfeeding and a rich vector of covariates and find that nursing is associated with a loss of 1.4 kilograms (three percent of body weight).3
The degree to which fertility is suppressed depends on how many times a day the mother breastfeeds and the intensity with which the child suckles (Rous, 2001). Furthermore, while the World Health Organization warns that breastfeeding reliably prevents pregnancy only during the first sixth months after delivery, many studies argue that this window is considerably longer in developing countries. For example, Weis (1993) and Thapa (1987) find that breastfeeding inhibits
3
Results available upon request. The analysis uses the third wave of the NFHS, the only wave with data on respondents’ weight. Covariates include a linear control for mother’s years of education, a linear and quadratic control for her age, a linear and quadratic control for her height, a linear and quadratic control of the child’s year-of-birth, and dummy variables for the state of residence, urban/rural, sex of the child, and child’s age in months.
fertility for 12 to 24 months in Bangladesh and Nepal, respectively. The longer duration of post-partum infertility in developing countries is likely related to mothers’ underlying malnutrition (Huffman et al., 1987; Yadava and Jain, 1998).
An important question for our model is whether mothers actually know about the contraceptive effects of breastfeeding. Typically, a nursing mother does not menstruate (a phenomenon known as “lactational amenorrhea”), which would presumably alert her to her temporary inability to conceive. The NFHS survey we use directly asks non-pregnant women the reason they are not using birth control. Excluding those who report they are currently trying to conceive, 34 percent cite breastfeeding as their reason for not using contraception.
2.2 Do women stop breastfeeding once they become pregnant again?
A related question is how subsequent conceptions or births affect the mother’s decision to continue breastfeeding the older child. According to the American Academy of Family Physicians (2008), there is no evidence that breastfeeding while pregnant is harmful to the fetus, but some speculate that it could increase the likelihood of miscarriages (Verd, Moll, and Villalonga, 2008). Similarly, the medical profession does not officially discourage breastfeeding two children at once (“tandem breastfeeding”), though there is some evidence that infant weight gain is slowed if a mother is simultaneously nursing an older sibling (Marquis et al., 2002).
Of course, the relevant question for this paper is whether mothers choose to stop breastfeeding after a conception or birth, for whatever reason. The mother’s decision might be driven not by the risks as perceived by the medical profession, but because, say, the opportunity cost of her time and energy rises after a subsequent pregnancy or birth. Moreover, meeting the caloric requirements of tandem breastfeeding is significantly more challenging in developing countries than it would be for the mothers in the above studies. Indeed, in the NFHS, the most common reason women cite when they stop nursing is “became pregnant,” (over 32 percent of respondents cite this reason).
3
Model
This section presents a simple dynamic model in which a mother’s decision to breastfeed and her future fertility depend on each other. The model predicts that breastfeeding rates will have several distinctive features with respect to children’s birth order, gender, and the gender composition of older siblings.
3.1 Overview
We assume that a mother has a per-period utility determined by the current number and sex composition of her children. After each birth, she must decide whether to breastfeed the child. If
she does not breastfeed, she will have another child in the next period. If she does breastfeed, she will not have another child in the next period. She makes this decision in order to maximize the infinite sum of discounted per-period expected utility.
Later in this section we discuss in greater detail many of the assumptions underlying this simple framework, but highlight one here: we assume that breastfeeding’s only function is contraception, and thus ignore any health benefits it might provide the child. Doing so allows us to demonstrate that our model can generate gender differentials in breastfeeding even when mothers value the health of daughters and sons equally. Introducing a health benefit would increase the magnitude of the male breastfeeding advantage our model generates but would not qualitatively change its other, more distinct predictions, for example regarding the interaction of gender and birth order.
3.2 Setup of the model
A mother’s utility depends on the number of children, n and of sons, s she has. Her period utility isu(n, s), and she has an infinite horizon with a discount rateβ. Time periods are denoted by t. In each period, a woman gives birth to either one child or no children.
A mother who gives birth in periodt decides whether to breastfeed the child, bt∈ {0,1}. We assume breastfeeding perfectly inhibits fertility in the subsequent period but has no ancillary costs or benefits. If bt = 1, then nt+1 =nt and st+1 =st. If bt = 0, thennt+1 =nt+ 1 and, since the next child is equally likely to be a boy or a girl, st+1=st+ 1 orst+1=st,each with probability 1/2. For the remainder of this section, we suppress the time subscript.
The breastfeeding decision, in essence, acts as a fertility stopping decision. For a mother who currently has n children of which s are sons, one option is not to breastfeed (continue having children), in which case she receivesu(n, s) this period and the discounted expected value function over subsequent periods. The value function in the next period is V(n+ 1, s) or V(n+ 1, s+ 1), with equal probability. The other option is to breastfeed (stop having children) and receiveu(n, s) in this and each of the infinite subsequent periods (giving a future lifetime discounted utility of u(n, s) +βu1(−n,sβ) = u1(−n,sβ)). We also assume that mothers have access to sterilization, but that even if they choose to get sterilized, they breastfeed their last child.4
The decision problem is therefore:
V(n, s) = max{Vb=1, Vb=0}= max u(n, s) 1−β , u(n, s) +β V(n+ 1, s) +V(n+ 1, s+ 1) 2 . 4The purpose of including the sterilization option is so that a mother can prevent further pregnancies after her last child is past the age of breastfeeding (one period in the model). Our assumption that mothers choose to breastfeed during this period instead of relying completely on sterilization does not seem unreasonable as mothers may still wish to pass on potential health benefits of breastfeeding. Alternatively, they may wish to delay their decision about sterilization given its irreversible effect and thus rely as long as possible on the contraceptive effects of breastfeeding.
We specify the period utility function as
u(n, s) =φf(n)−c(n)
| {z }
≡q(n)
+λg(s).
The term q(n) ≡ φf(n)−c(n) captures the net benefits (benefits, φf(n), minus costs, c(n)) of having n children, with φ > 0 parameterizing the demand for children. We assume that f00 < 0 andc00>0,soq00<0, for alln, and further that limn→∞q0(n) =−∞. In other words, with respect to the number of children n, themarginal net benefit of an additional child is strictly decreasing (and falls without bound) and thetotal net benefit displays an inverted-u shape.
The termλg(s) represents the additional utility from sons, with λ≥ 0 measuring the degree of son preference. We assume g0 >0 andg00 <0 for all s. Note that for convenience we consider smooth f, c, and g defined over R+, despite the fact that in a woman’s choice set, n and sonly
take on integer values.
A useful quantity to define is the value ofnup to which a mother would choose to have another child regardless of her son preference or the sex composition of her children. We call this quantity
b
n.
Definition. Let bn= max{n|q(n+ 1)−q(n)≥0}.
Son preference factors into the breastfeeding decision once n >bn.Intuitively, mothers unambigu-ously gain from having up to bnchildren, after which point they weigh the net cost of having more children (which grows unboundedly) against the value of trying for more sons.5
3.3 Predictions of the model
This subsection presents several predictions of the model regarding the dependence of breast-feeding on a child’s birth order, gender, and the sex composition of his older siblings.
Proposition 1. Breastfeeding is increasing in birth order.
Proof. A mother who stops at n children will have breastfed her nth child but not breastfed her first n−1 children. This follows from the equation of motion for n. Once a mother chooses to stop having children, she will not resume having children: suppose she did; she could increase her lifetime utility by shifting her childbearing earlier given her positive discount rate.
In addition to depending (weakly) on birth order, breastfeeding also depends on gender. 5Note that we are interested in the case where a mother wants to have at least one child, or
b
n >0 (equivalently, q(1)−q(0)>0). Lemma 1 in the Appendix shows that under this assumption there exists a uniquen, and that itb always lies in (nmax−1, nmax) wherenmax is then that maximizesq. Also note that even a mother with all sons might continue pastbnbecause of the marginal benefit of more sonsλ[g(bn+ 1)−g(bn)], but every mother continues up to at leastn.b
Proposition 2. At any birth order, a child is more likely to be breastfed if
(i) the child is a son rather than a daughter (holding fixed the gender of older siblings); or (ii) more of his or her older siblings are male (holding fixed the child’s gender).
Proof.
(i) By Lemma 2 (in the Appendix), a mother will breastfeed if and only ifu(n, s)≥ u(n+1,s)+u2(n+1,s+1). Keeping the terms in the utility function that depend on s, breastfeeding is increasing in g(s)−g(s+ 1), which is increasing inssinceg00<0. Holding the sex of the firstn−1 children fixed, s is higher when the nth child is a boy, so a son is more likely to be breastfed than a daughter. For sufficiently large λ, there exist integer values of nand s≤nfor which the mother will choose to breastfeed a son but not a daughter.
(ii) From the proof to part (i), breastfeeding is increasing in s. Holding the sex of the nth child fixed, s is increasing in the number of boys among the first n−1 children, so an nth child
with more brothers among his or her siblings is more likely to be breastfed. For sufficiently largeλ, there exist integer values ofnands≤nfor which the proposition holds strictly.
In the model, breastfeeding does not enter the utility function through its effects on the child who is nursed, so there is no difference in how much the mother values breastfeeding her sons versus her daughters per se. Instead, the breastfeeding gender gap is caused by fertility stopping preferences. Moreover, through this mechanism, not just the gender of the child, but also the gender of his or her older siblings affects breastfeeding.
Perhaps the least obvious prediction is that the gender gap in breastfeeding depends non-monotonically on birth order.
Proposition 3. The largest gap in breastfeeding of boys versus girls is at middle birth order. In other words, the gap is increasing with birth order for sufficiently low birth order, and decreasing in birth order for sufficiently high birth order.
Proof. At sufficiently low birth order, breastfeeding is the same for sons and daughters: Lemma 1 establishes that mothers never breastfeed any child before their bn
th. At sufficiently high birth
order, breastfeeding is again the same for sons and daughters: by Lemma 3, mothers eventually breastfeed because they eventually want to stop having children. By Proposition 2(i), the only form of gender gap that can obtain is when a mother would breastfeed a son but not a daughter. If a mother’s preferences are such that this would never obtain, the proposition holds trivially. If there are some birth orders (and gender compositions of older siblings) for which a mother would
breastfeed a son but not a daughter, the gender gap increases when the first such child is born and decreases when the last such child is born.
The intuition behind this result is that at low n, mothers want to continue having children regardless of the sex composition of their existing children sinceφf(n)−c(n) is still increasing in n. Therefore they will breastfeed neither sons nor daughters. At high enough n, the net cost of increasingnbecomes large enough that it outweighs any benefit of having another son. A mother will breastfeed both a son or a daughter in this case. When weighing the costs of higher quantity with the benefit of having (in expectation) more sons at intermediate values of n, however, a mother may want to stop having children if and only if her nth child is male, and will therefore breastfeed a son but not a daughter.
3.4 Predictions regarding “ideal family size”
The model also predicts that breastfeeding patterns can change abruptly when n reaches n.b While a mother’s “ideal” quantity of children is ill-defined in the presence of son preference, one can think of bnas one measure of the mother’s preferred quantity: it is the quantity at which she would want to stop having children if she had no son preference (λ= 0). Empirically, we will use survey questions on ideal family size to test the predictions below, and we refer to bn as “ideal
family size.”
Proposition 4.
(i) Breastfeeding is constant for birth order below the ideal family size and can strictly increase in birth order only after the ideal family size has been reached.
(ii) There is no gender gap in breastfeeding for birth order below the ideal family size. The gender gap in breastfeeding only arises after the ideal family size has been reached.
Proof.
(i) A mother will not breastfeed forn <bn, by Lemma 1. Proposition 1 establishes the remainder of the proof. The mother will begin breastfeeding at somen≥bn, the exact value depending on the gender composition of her children and the extent of her son preference.
(ii) The proof to part (i) establishes that neither boys nor girls are breastfed for n <nb, so there is no gender gap. Proposition 2(i) establishes the possibility of a gender gap at n≥nb.
3.5 Discussion
We close this section by revisiting some of the model’s assumptions, beginning with our decision to model breastfeeding as a binary choice. The model can easily be relabeled so that the decision
is between short and long periods of nursing, to reflect that fact that almost all children in our sample are breastfed initially, but the binary nature of the model still deserves further attention. First, modeling the decision as binary makes the predictions hold weakly for an individual mother. For example, a mother with negligible son preference who wants three children will breastfeed her third child but not the first two (or, equivalently, {short, short, long}). In line with Proposition 1, her breastfeeding choice is weakly increasing in birth order for all birth or-ders; it is constant between birth order one and two, and then strictly increases from birth order two to three. Empirically, we examine a large population of mothers, and given heterogeneity in ideal family size and son preference, aggregating over a population should smooth out the discrete fertility-stopping decisions predicted by the model. In addition, patterns for a particular mother depend on the realized sex composition of her previous children, whereas for a population, the population-average sex composition is known. Thus, we conjecture that the results would hold strictly (e.g., breastfeeding is strictly increasing in birth order) at the population level for most well-behaved joint distributions of φ(demand for children) andλ(son preference).
Second, the binary specification does not take into account mothers’ using breastfeeding to space births as opposed to merely prevent them. The essential assumption to incorporate birth-spacing into our framework is that mothers who want to continue having children might space their births to an extent, but they will still breastfeed less than mothers who want to stop having children entirely. This assumption follows naturally from a model (such as ours) where a mother receives flow utility from her children; she will begin receiving the positive flow of utility sooner if she has her children sooner.6
It is also worth reviewing the assumption that breastfeeding perfectly prevents conception, even though the studies cited in the previous section indicate it merely decreases fecundity. Allowing mothers to conceive while breastfeeding would only reinforce the model’s predictions. Recall from the previous section that mothers who become pregnant while breastfeeding tend to wean the current child, which suggests a causal relationshipfromfuture fertilitytobreastfeeding in addition to the causal relationship we model from breastfeeding to future fertility. Both mechanisms predict a negative relationship between a mother’s tendency to breastfeed and her desired future fertility, and we make the choice to model only one in the interest of simplicity.7
6Stepping outside the model, fecundity declines with age, so delaying childbearing also carries the risk that the mother becomes less able to conceive.
7For example, consider a modification of the model such that a mother chooses whether to breastfeed her child both in the period in which the child is born and in the next period. Assume a mother is constrained to breastfeed only one child at a time, and she prefers to breastfeed a younger child (which follows directly if she perceives the health benefits for each child as concave in breastfeeding duration). Then a mother will breastfeed her child in the second period of his or her life if and only if the child does not have a younger sibling, or equivalently if and only if the child was breastfed in the first period. Thus, the two channels reinforce each other.
Even with these simplifications, the model makes several distinct empirical predictions. Other theories might be able to generate the prediction that breastfeeding depends on birth order (e.g., learning-by-doing models) or gender (e.g., mothers place more value on the health of sons). How-ever, alternative hypotheses would also have to explain why breastfeeding depends on older siblings’ genderand why the breastfeeding gender gap has an inverted-ushape with respect to birth order. Similarly, alternative hypotheses would have to explain why breastfeeding sharply increases once mothers reach their ideal family size, and why the male breastfeeding advantage also increases just at this point. Thus, we feel that evidence of these effects would collectively provide strong support for our claim that women’s preferences over future fertility affect the decision of when to wean their children.
4
Data
Our empirical analysis uses the 1992, 1998 and 2005 waves of the National Fertility and Health Survey (NFHS) of India, a repeated cross-sectional data set based on the Demographic and Health Survey. The NFHS surveys a representative sample of ever-married women ages 15 to 49 across India.
The main advantage of the data set for our purposes is that it records the number of months all children under three, four or five years (depending on the wave) were nursed. Additionally, as basic demographic information is recorded for every child born to a survey mother, we can calculate variables such as birth order and the sex composition of siblings for each child. The survey also includes a variety of information on contraception, desired fertility, and child health, as well as standard demographic and household characteristics.
We make several sampling restrictions. First, our observations are at the child level, so we include only women who have given birth to at least one child. Second, we exclude observations with missing values for duration of nursing, which restricts the survey to relatively recent births since the survey does not collect retrospective breastfeeding information for older children; the data were collected for children up to age four for the first wave, age three for the second wave, and age five for the third wave. (Because the duration of breastfeeding is censored at 36 months in the second wave, we top-code the variable at 36 months for all waves.) In addition, for about one percent of children whose breastfeeding information was actually solicited, the data are missing or were deemed “inconsistent” by the surveyors. Third, we exclude mothers who have eight or more children (the 95th percentile for this variable) to reduce composition bias from mothers with unusually large family size. Fourth, we exclude multiple births (e.g., twins) since their birth order is less well-defined and how long they are breastfed might systematically differ by virtue of
their not being a singleton. Finally, for the breastfeeding analysis, we exclude children who have died, as otherwise the nursing period would be censored in such a manner that does not reflect mothers’ preferences regarding breastfeeding; this restriction results in a loss of about four percent of remaining observations. Our final sample includes just over 110,000 observations.8
Summary statistics for the sample used in the breastfeeding analysis are presented in Table 1. The mean duration of breastfeeding, adjusted for censoring due to children still being breastfed, is 22.8 months. The mean value of birth order is 2.6. We observe women who may not have completed their fertility, so we cannot calculate total fertility in our data. However, based on Census data, total fertility conditional on having at least one child (the population from which our sample is drawn) is roughly four children per mother for the cohorts in our data.9
Although the data set covers a large sample of children and includes detailed breastfeeding and contraceptive information, it also presents several challenges with respect to our hypothesis-testing. For example, the breastfeeding variable is only recorded for relatively young children, so we can only observe breastfeeding duration for at most one or two children per mother.10 As such, comparing breastfeeding across siblings with a mother-fixed-effect model is infeasible, and our analysis is subject to some of the standard criticisms of cross-sectional estimation, which we attempt to address in the next two sections.
5
Empirical analysis: Breastfeeding as a function of birth order
5.1 Estimation strategy
Our first prediction relates breastfeeding duration to a child’s birth order. We test the hypoth-esis that breastfeeding is a positive function of the mother’s desire to inhibit her future fertility, which increases with birth order.
We begin by imposing as little structure as possible, allowing birth order to enter non-parametrically as a vector of dummy variables. We estimate the following OLS model:
Breastf eedi =
X
k
βk·1(BirthOrderi =k) +Xi·γ+ai+εi. (1) Breastf eedi is the number of months a mother reports having breastfed child i. The term ai is a vector of age-in-months fixed effects up to 36 months (the maximum value of the outcome)
8
For the analysis of child mortality, we obviously do not condition on being alive and are not limited to births for which breastfeeding data are available; we describe the construction of the mortality sample in section 7.
9
In the 1991 Census, mothers age 50-54 had a mean total fertility of 4.8. We assume fertility is falling at 1% a year. Disaggregated fertility data for the 2001 Census are not yet publicly available.
10A further problem that invalidates a mother-fixed-effect model is that there is a mechanical correlation between breastfeeding and inclusion in the sample: For a mother to have two children young enough for their breastfeeding data to be available, she must not have breastfed the older one for very long.
that mechanically accounts for the fact that recently-born children will appear to have shorter breastfeeding spells due to right-censoring. X is a vector of covariates, and εis an error term.
The OLS model has the advantage that the coefficients are easy to interpret, but the right-censoring must be mechanically corrected with age-in-months fixed effects. We therefore also esti-mate the following proportional hazard model, which accounts for the fact that many children in our regression sample are still being nursed, imposes no conditions on the baseline hazard function, and models independent variables as having a proportional effect on the hazard rate:
hi(t) =h0(t)∗exp X k βk·1(BirthOrderi=k) +Xi·γ+εi ! . (2)
We show the results from these and related estimations in the following graphs and tables.
5.2 Results on birth order
Figure 1 plots birth order on the horizontal axis and the estimated coefficients on the birth-order dummy variables on the vertical axis. We show the coefficients from both an OLS and a hazard model. To make the coefficients comparable, the hazard coefficients are for “survival” rather than “failure”; a larger coefficient implies that the variable is associated with a longer duration of breastfeeding. As the figure shows, the pattern of coefficients is similar regardless of which specification is used. The figure also shows the distribution of birth order in the data.
The first three columns of Table 2 report regression results that summarize the effect of birth order on breastfeeding duration. The first column shows the linear coefficient on birth order from an OLS regression with no other covariates, which suggests that a one-unit increase in birth order is associated with a 0.46 month increase in breastfeeding duration.
Higher birth-order children are born to older mothers, are born more recently, and belong to a larger family. If breastfeeding has been trending over time or if mothers with higher fertility differ in their propensity to breastfeed, for example because they are less attached to the labor force, then the coefficient on birth order could suffer from omitted variable bias. Thus, col. (2) includes a linear control for mother’s years of education, a linear and quadratic control for her age, a linear and quadratic control for the child’s birth year, and dummy variables for the survey wave, state of residence, urban/rural, and sex of the child. These covariates directly address mother’s age and time trends as confounding factors and also include proxies for likely total fertility, such as mother’s education. With the controls added, the estimated coefficient falls to 0.21. Col. (3) is the hazard estimation analogue to col. (2) and suggests that a one-unit increase in birth order is associated with a six-percent decrease in the probability of being weaned in any given month.
shrinks the coefficient on birth order by roughly fifty percent suggests that birth order is correlated with other variables that predict breastfeeding duration. While the rich set of controls in col. (2) partially addresses this problem, it is impossible to completely eliminate concern about omitted variable bias for the birth-order results. Any variable that has a roughly linear relationship with birth order could be a potential confounding variable, and there are likely to be many. And even with the most comprehensive controls, it is difficult to account for the fact that, mechanically, birth order is partly determined by family size. For reasons previously described, including mother fixed effects is not a viable option using our data.
Moreover, even a causal effect of birth order on breastfeeding would not be definitive evi-dence of our hypothesis that future fertility determines breastfeeding choices. With respect to the relationship between breastfeeding duration and birth order, our hypothesis is observationally equivalent to a “learning by doing” model in which a mother’s cost of breastfeeding declines as she gains experience with each subsequent birth.
For all of these reasons, while it is reassuring that the basic birth-order results fit the model’s prediction, we consider the results presented in the remainder of this section and in the next section to be stronger tests of the model.
5.3 Results on “distance from ideal family size”
Our model makes predictions regarding not only a child’s birth order but his birth orderrelative to his mother’s “ideal” family size. Specifically, the model implies that breastfeeding is constant for birth order below the ideal family size and increases in birth order only after the ideal family size has been reached (Proposition 4(i)). Fortunately, our data set includes each mother’s self-reported “ideal” family size.11From this measure we generate a variable that measures the current distance from the mother’s ideal family size: ∆Idealij =BirthOrderij−Idealj (whereBirthOrderij is the birth order of mother j’s ith child andIdealj is mother j’s ideal family size). This variable allows us to test this prediction of the model.
Before describing our estimation strategy and results, we highlight some potential problems with the “ideal” family size measure. First, the concept of ideal family size is not well-defined without reference to sex composition. For example, a mother with no children might say her ideal family size is two, thinking her first two children would be boys; if she knew she would have two girls first, then her ideal family size might be larger.
Second, to avoid cognitive dissonance, mothers might self-report their “ideal” fertility prefer-ence to match their actual fertility outcome, and thus the variable might not reflect their
pref-11
The survey question is, “If you could go back to the time you did not have any children and could choose exactly the number of children to have your whole life, how many would that be?”
erences at the time they gave birth and made breastfeeding choices (Pritchett, 1994). However, we suspect that such ex-post rationalization is limited. Table 1 indicates that actual fertility sys-tematically exceeds self-reported ideal fertility; ideal family size is on average 2.7, but based on the Census, total fertility is about four for these cohorts. Furthermore, the histogram in Figure 2 shows that many children are born to mothers who have already reached their ideal family size (i.e., ∆Ideal >0).12
Figure 2 plots the estimated coefficients on the ∆Ideal dummy variables from a regression analogous to equation (1) that replaces BirthOrder with ∆Ideal. The point estimates display a similar increasing pattern as those for birth order, which is not surprising as ∆Idealis increasing in birth order (though not merely a linear transformation, asIdealj varies for each mother). The figure also suggests there is a jump in the level of breastfeeding once ideal family size is reached (∆Ideal= 0), as predicted by the model.13
The last four columns of Table 2 show results from regressions relating breastfeeding duration and ∆Ideal. We specify the ∆Idealeffect as a level increase once mothers reach their ideal fertility. Col. (4) of Table 2 shows the coefficient on an indicator variable for ∆Ideal ≥0 when no other covariates are included; once mothers reach their ideal fertility they breastfeed subsequent children an extra 1.07 months. Adding covariates in col. (5) yields a coefficient of 0.88 (an 18 percent drop). That adding covariates decreases the ∆Ideal ≥ 0 coefficient much less than the birth-order coefficient suggests that the ∆Idealregressions are less vulnerable to composition bias since ∆Ideal implicitly accounts for heterogeneity in fertility preferences.
The more exacting specification, though, is to test for a discrete increase in breastfeeding at ∆Ideal = 0 while allowing for an overall linear effect of ∆Ideal. We estimate the following equation, where again the variable of interest is the indicator for ∆Ideal≥0:
Breastf eedij =δ·1(∆Idealij ≥0) +λ·∆Idealij+Xi·γ+ai+εij. (3) Col. (5) shows the results without our set of covariates included. Once mothers reach their ideal fer-tility, the duration they breastfeed subsequent children increases by 0.40 months. With covariates included, the effect now becomes stronger: Above and beyond an overall linear effect of ∆Ideal, breastfeeding duration increases by 0.58 months once a mother reaches her ideal family size.
12Two thirds of mothers in our sample age 35 years or older have more children than their ideal number; these older mothers are more likely to have completed their fertility, but even they might continue having more children. 13To explore more rigorously the discrete jump in breastfeeding when a mother reaches her desired number of children, we create a series of dummy variables corresponding to the conditions ∆Ideal≥xand then run separate regressions of breastfeeding on each dummy variable (as well the standard covariates). Indeed, the regression using the variable ∆Ideal≥0 yields the largestt-statistic andR2. Results available upon request.
5.4 Discussion
There appears to be a strong positive correlation between breastfeeding and birth order, con-sistent with the model’s prediction. As the average mother has about four children, the last child would be breastfed about 0.6 months longer than his oldest sibling. Similarly, children born once their mother has reached her “ideal” family size are breastfed 0.6 months longer than older sib-lings. While this effect size might not seem large, a breastfeeding differential of this magnitude has important consequences for child mortality, as we show in Section 7.
While we have already discussed potential biases in our estimates, especially those regarding birth order, several pieces of evidence from this section point to a causal relationship between desired fertility and breastfeeding. First, while composition bias may indeed affect the coefficient on birth order in Table 2, such bias alone cannot explain our finding in Figure 1 that breastfeeding increases between birth order one and two, since almost all mothers in India have at least two children. Moreover, the marked increase in breastfeeding just at the point when mothers reach their ideal family size suggests that they take into account their fertility preferences when deciding when to wean their children.
We now turn to testing our model’s distinctive predictions regarding how the child’s sex, the sex composition of existing children, birth order, and desired fertility interact to predict breastfeeding duration.
6
Breastfeeding as a function of gender and birth order
There are many reasons a mother may decide to breastfeed sons longer than daughters. Daugh-ters might simply be harder to nurse, or sons might be harder to wean. Alternatively, parents in India may choose to allocate more resources to sons (Das Gupta, 1987; Pande, 2003; Mishra, Roy, and Retherford, 2004; Oster, 2009). If mothers perceive breastfeeding as superior to alternatives such as infant formula or solid food, then by this logic they will nurse sons longer.
Our model offers a different explanation, namely that a preference for having a future son causes a gender gap in breastfeeding thecurrent child.14Demand for an additional child is higher after the birth of a girl and thus mothers wean daughters sooner in the hopes of conceiving again (Proposition 2(i)). However, this prediction alone does not allow us to distinguish our hypothesis from the explanation that mothers value the health of sons more than daughters and nurse sons more in the belief that “breast milk is best,” or that sons simply “take to the breast” more easily than do daughters. In the next subsection, we discuss the predictions of the model that allow
14
Researchers find that in India, at any birth order, a couple is more likely to stop having children if they have just had a son (Das, 1987; Mutharayappa, Choe, Arnold, and Roy, 1997). Birth intervals are also longer after the birth of a son (Arnold, Choe, and Roy, 1998; Retherford and Roy, 2003).
greater separation of our hypothesis from these alternatives.
6.1 Testing our hypothesis
First, since parents’ preference for another son depends on the gender composition of all pre-vious children and not just the last one, there should be a separate effect on breastfeeding of variables such as “already has a son” and “percent of children that are male” (Proposition 2(ii)). While these variables are highly correlated with the sex of the current child, they should exhibit a separate effect in our model but not in other models of breastfeeding in which, say, mothers simply value the health of sons more than the health of daughters.
Second, the model suggests that the effect of a child’s gender and other sex-composition vari-ables are strongest at intermediate birth order (Proposition 3). At very low (high) birth order, mothers want to continue (stop) having children regardless of sex composition. Any confounding variable would have to cause a similar, non-monotonic gender differential with respect to birth order.
Third, our model predicts that the male breastfeeding advantage should become most pro-nounced once mothers reach their “ideal” family size (Proposition 4(ii)). Again, potential con-founding variables would have to conform to this specific pattern.
6.2 Results on breastfeeding as a function of gender and sex composition Figure 3 plots breastfeeding duration (the survival function) separately for boys and girls. The graphs do not hold any other variables constant but do account for censoring of the breastfeeding duration variable due to some mothers still breastfeeding at the time of the survey. Though the heaping of observations at multiples of six months makes it difficult to see the differences across gender at those specific points, one can compare percentiles for which both distributions are far from a heaping point.15 For example, the 30th percentile for boys is about one month more than that for girls. At the 70th percentile the difference is about 2.5 months.
Table 3 shows the effect of the child’s sex and the sex composition of existing children on breastfeeding duration. The first column indicates that sons receive an additional 0.37 months of breastfeeding relative to daughters. This effect barely moves after adding our standard set of covariates from Table 2 plus birth-order fixed effects in the second column. Col. (3), the hazard-model analogue of col. (2), suggests that sons have a ten percent lower probability of being weaned in any given month relative to daughters.
15Some of the observed age-heaping may reflect true discontinuities in behavior and not merely rounding error. For example, the Qur’an calls for children to be nursed for 24 months (Yurdakok, 1988). Indeed, we find that Muslim mothers are more likely to report having weaned their child at exactly 24 months. When we exclude Muslims mothers from our regression sample, all of our relevant effect sizes grow, which is not surprising since following this religious dictate should make these mothers less responsive to the incentives described in our model.
In the next two columns we examine whether the sex composition of siblings has an independent effect on breastfeeding even after accounting for the sex of the current child. Col. (4) shows that a mother already having at least one son increases the current child’s breastfeeding duration by 0.28 months. In other words, the breastfeeding gap between two girls, one of whom has an older brother and the other of whom does not, is almost as large as the breastfeeding gap between boys and girls. Similarly, mothers breastfeed a child longer when the male share of her other children is high. Thus, there is strong evidence that the gender composition ofpast births affects the breastfeeding of the current child—a pattern not easily predicted by theories that assume mothers prefer to breastfeed sons or that sons “take to the breast” better than daughters.
Finally, the last column of Table 3 re-estimates the specification in col. (2), allowing the male coefficient to vary by survey wave. The main effect is the gender differential in breastfeeding for the most recent wave of data, collected in 2005; the coefficient of 0.46 is about 18 percent higher than the pooled gender differential seen in col. (2). The male-wave interactions are imprecisely estimated but suggest that the gender gap in breastfeeding has been increasing over time from 0.31 months in the first wave to 0.39 months in the second wave to 0.46 months most recently. This increase over time is consistent with previous evidence that India’s recent fertility decline has intensified the country’s sex bias (Das Gupta and Bhat, 1997).16,17
6.3 Gender effects as a function of birth order
We now examine how the gender differences seen in the previous subsection vary with birth order. Recall that our model predicts that the gender effects are small for both high and low birth order. Moreover, for the population as a whole, the peak effect of these variables should occur between the average “ideal family size” (around three) and average total completed fertility (around four).18 Therefore, our model not only predicts that the gender effect takes an inverted-u shape with respect to birth order but also specifies the birth-order interval at which it peaks.
In order to investigate the effect of birth order on the gender coefficients in a flexible manner, 16A decrease in fertility will intensify sex bias if the desired number of children falls more rapidly than the desired number of sons. See, for example, Das Gupta and Bhat (1997) and Arnold, Kishor, and Roy (2002). This effect seems to outweigh the fact that as sex selective abortion becomes more available, girls should be born more often into families with less son preference.
17
Our other empirical results, those presented in the previous section and in the remainder of the paper, are also similar when estimated separately by survey wave.
18
Recall the two mechanisms underlying the breastfeeding-fertility relationship. If the relationship depended en-tirely on women using breastfeeding as contraception, then the population-average effect would peak at the average value for “ideal” fertility. If the relationship depends entirely on subsequent pregnancies causing women to stop nursing the current child, then the effect would peak at the average completed fertility. As both channels appear to operate in our data, we predict the effect to peak somewhere between these two values.
we estimate the following equation:
Breastf eedi = α·M ale+
X k βk·1(BirthOrderi=k) +X k δk·M ale×1(BirthOrderi=k) +ai+Xi·γ+εi. (4) The key addition is the vector of M ale×BirthOrder dummy variables, which allows each com-bination of gender and birth order to have its own fixed effect.
Figure 4 plots the estimated breastfeeding durations by birth order and gender from the above estimation without the additional control variablesX. Sons are breastfed more than girls at every birth order, but the difference is not constant across birth order. The difference is increasing until birth order four, and then decreasing after that, in line with the model’s predictions. The male-female difference by birth order is also plotted to show the inverted-u shape more clearly.
Based on the evidence in Figure 4, we specify the gender effect parametrically as a quadratic function of birth order. These regression results are reported in the first three columns of Table 4. Col. (1) shows the OLS estimate, excluding the control variables X. The coefficients for the quadratic terms suggest that sons’ breastfeeding advantage peaks when birth order equals roughly 4.1. Cols. (2) and (3) show that this peak is robust to adding in the control variables or using a hazard model.
Figure 5 plots breastfeeding duration by gender, but this time against ∆Ideal, the distance between the birth order of the current child and the mother’s “ideal family size” (recall that the variable ∆Idealij = BirthOrderij −Idealj, where BirthOrderij is the birth order of mother
j’s ith child and Idealj is mother j’s ideal family size). The male-female difference appears to increase just as mothers reach their “ideal” family size and then slowly narrows. Indeed, allowing a level increase once ∆Ideal reaches zero is the specification most favored by the data. (Results available upon request.) We read the evidence in Figure 5 as consistent with our prediction that gender preferences will have the largest effect on breastfeeding when the mother’s decision to have another child is most marginal.
The final two columns of Table 4 present the regression analogue to Figure 5. Even after controlling for our standard covariates and allowing for an interaction of M aleand ∆Ideal, there appears to be a discrete jump in the male breastfeeding advantage just when mothers reach their “ideal” family size. This advantage jumps by 0.49 months when a mother reaches her ideal family size. The effect size is unchanged when covariates are added to the regression.
6.4 Relating our results to fertility outcomes
Our model predicts a negative relationship between breastfeeding and subsequent fertility, both because mothers who want to conceive again wish to limit nursing-induced postpartum amenorrhea and because mothers who conceive again generally stop breastfeeding. In this section, we verify these channels by testing whether the same variables that predict breastfeeding also predict fertility-stopping.
Table 5 shows results when “no younger sibling” serves as the dependent variable and the right-hand side takes on many of the specifications explored in this section and the previous one. For comparability with the previous results, we use the same sample as before. Col. (1) shows that high-birth-order children are more likely not to have a younger sibling at the time of the survey, and col. (2) shows an increase in this probability once mothers reach their “ideal” family size.
The next columns focus on the child’s gender and the sex composition of the child’s siblings. Being male increases the probability of not having a younger sibling, as does having at least one older brother—the same pattern as when breastfeeding served as the dependent variable. Finally, the interactions between gender and birth order echo the patterns we found for breastfeeding. A son’s increased likelihood of being the youngest child peaks when birth order equals about 3.5. In addition, the amount by which boys are more likely than girls to be the youngest child is greater for mothers who have reached or surpassed their ideal family size.
6.5 Heterogeneity in son preference
We next test whether gender patterns in breastfeeding vary with the intensity of the mother’s son preference. Our first measure of son preference is the male-to-female ratio of births in the respondent’s state (using data from 2001, the most recent census year), which has a sample average of 1.07. Given sex-selective abortion, this statistic should be positively correlated with regional differences in son preference, and indeed we find that it is systematically lower in south India, where son preference is recognized to be less intense.
Table 6, col. (1) shows that the son advantage in breastfeeding is significantly larger in states with stronger son preference. The estimates imply that in states such as Kerala with sex ratios near one, mothers breastfeed sons only 0.2 months longer than daughters; in places such as Haryana or Punjab where the sex ratio is about 1.16, boys are breastfed 0.6 months longer. This basic result is consistent with both a stop-after-a-boy fertility preference and with parents caring more about sons’ health.
We next examine how the breastfeeding gap increases discretely when the mother has reached her ideal family size, which is a more distinctive prediction of son-biased stopping rules. As seen in
col. (2), our predicted patterns are stronger in states where the sex ratio is higher. That is, in states where having sons is especially valued, the mother’s decision to continue having children beyond her ideal number, and thus her decision about how long to breastfeed, depends more heavily on the gender of her children.
Another measure of son preference is how many sons the mother says she ideally would like to have. The survey asked the mother to break down her ideal number of children into her ideal number of sons (sample mean of 1.4), ideal number of daughters (mean of 1.0) and ideal number for which she was indifferent about their gender (mean of 0.3). The first prediction we test is that, just as a mother is more likely to stop having children once she reaches her ideal family size, she is more likely to stop once she reaches her ideal number of sons. We construct a variable ∆IdealSons that is analogous to ∆Ideal; ∆IdealSons is the mother’s current number of sons minus her ideal number of sons. (For each child-level observation, the mother’s current number of sons is the number she has at the time of the child’s birth, and in the case of a boy, it is inclusive of the child himself.) As seen in col. (3), the duration of breastfeeding increases by 0.33 months once the mother reaches her ideal number of sons, controlling for a linear effect of IdealSons. Once a mother has reached both her ideal number of sons and her ideal family size, her children are breastfed 0.83 months longer.
We next test a second prediction about IdealSons. The scenario in which a child’s gender is most pivotal to the breastfeeding-cum-fertility decision is when the mother, by giving birth to a son, has just reached her ideal number of sons. Because the child was male, her son preference is satisfied; had she given birth to a girl instead, her son preference would not have been met. Thus, empirically, the male advantage in breastfeeding should be most pronounced when the mother has exactly reached her ideal number of sons, compared to when she has either fewer or more than her ideal number of sons. As seen in col. (4), we find precisely this pattern: relative to all other values of ∆IdealSons, the case where ∆IdealSons = 0 is when boys are breastfed most, relative to girls.19 In short, we find strong evidence that the son advantage in breastfeeding varies as predicted with heterogeneity across regions in son preference, and breastfeeding duration varies across mothers who differ in their ideal number of sons, with quite specific predictions born out in the data.
6.6 Decomposing the male breastfeeding advantage
Thus far, we have provided a variety of evidence that boys are breastfed longer than girls, and that the gender interactions with birth order and ideal family size are consistent with distinct
19
The results in columns (3) and (4) are unchanged when we control for a set of fixed effects for the number of sons the mother has and the identification comes from variation across mothers in their ideal number of sons.
predictions of our model in which mothers decide whether to breastfeed based on their future fertil-ity. These predictions allow us to separate our mechanism from the more standard son-preference explanation that mothers simply give fewer resources—including breast milk—to daughters.
However, we have yet to calculate just what share of the male breastfeeding advantage our hypothesis explains. Overall, there is a 0.39-month male breastfeeding differential (Table 3, col. 2), which could be driven by our hypothesis, the more standard “feed the boys, starve the girls” explanation, or something else. One approach to gauging the importance of our mechanism is to assume that it does not enter until mothers reach their ideal family size. This estimate is provided by the coefficient on M ale∗∆Ideal ≥ 0 in col. (5) of Table 4, which is 0.495. As the mean of ∆Ideal ≥ 0 is 0.55, we estimate that our mechanism explains 0.272 months (0.495*0.55) or 70 percent (0.27/0.39) of the male breastfeeding advantage. Another approach is to estimate the “feed the boys” effect and all other explanations besides ours as the male coefficient conditional on the mother’s total number of children and total number of sons; in our model, these variables (nands) fully determine breastfeeding. In unreported results, we find the male coefficient is then 0.146, suggesting that our hypothesized mechanism explains 62 percent (1 - 0.15/0.39) of the breastfeeding gender gap.
As further evidence that the conventional son-preference story is unlikely to explain our results, we find no evidence that vaccinations or doctor visits exhibit the same distinct patterns we find for breastfeeding. Table 7 shows the results for our main specifications when “child received at least one vaccination” serves as the dependent variable. While there is indeed a strong male advantage, none of the other patterns found for breastfeeding hold. Col. (1) shows vaccinations are in fact decreasing with parity. Col. (3) shows that the sex of older siblings does not affect vaccinations, and col. (5) shows that the male advantage seems to actually decrease once the mother reaches her ideal family size. We find similar results (available upon request) using “total number of vaccinations” or “received medical attention conditional on having a fever” as the dependent variable.
6.7 Discussion
In this section we have presented evidence in support of the more demanding predictions of our hypothesis that future fertility affects breastfeeding duration. First, not only are boys breastfed for longer periods (a result consistent with mothers valuing boys’ health more than girls’ health or simply wanting to be more loving toward sons), but the gender of older children affects the breastfeeding duration of thecurrent child. Second, this gender effect takes an inverted-ushape: it is smallest at low birth order when a mother wants to continue having children (limit breastfeeding) regardless of her children’s gender and at high birth order when she wants to stop having children (prolong breastfeeding) regardless of gender. Third, the peak gender effect occurs at a birth order
between average ideal family size in our sample and realized family size in the Indian census; at these values of birth order, mothers’ decisions to conceive again are the most marginal. Fourth, the male breastfeeding advantage displays a discrete increase once mothers reach their self-reported ideal family size.
We also find that future fertility exhibits the same patterns, and that the results vary with measures of son preference. Finally, we find that other inputs to child health such as vaccinations or doctor visits also show a son advantage, but they exhibit none of the more distinct patterns we find for breastfeeding that are predicted by our model. Parents wanting to provide more health inputs to sons no doubt plays a key role in children’s health outcomes in India. However, we estimate it accounts for only about a third of the gender differential in breastfeeding.
7
Breastfeeding and child health patterns in India
In this section, we examine the implications of our results for infant and child health patterns in India. We first discuss the medical evidence on the health benefits of breastfeeding. We then review the patterns for child health and mortality across gender and birth order documented by past researchers to determine whether they appear plausibly related to the patterns we find for breastfeeding. Finally, we use our NFHS sample to directly test whether child survival exhibits the same relationships with gender, birth order and ∆Ideal as does breastfeeding duration.
7.1 Breastfeeding and infant and child health in developing countries
Medical and public health researchers have suggested several mechanisms by which breastfeed-ing promotes health for infants and young children in developbreastfeed-ing countries. First, human milk has immunological benefits; for example, it contains glycans that are believed to play an anti-infective role in the gastrointestinal tract (Morrow et al., 2005). Second, breastfeeding allows infants to avoid contaminated food and water, a mechanism that can play an especially important role in environments with poor sanitation (Habicht, DaVanzo, and Butz, 1988).
Much of the empirical work has focused on breastfeeding’s effect on mortality. Victoria et al. (1987) and Betran et al. (1987) find that breastfeeding is associated with lower rates of infant mortality from diarrheal disease and acute respiratory infection in Latin America, and Chen, Yu, and Li (1988) find similar results in China. Retherford et al. (1989) find that controlling for breastfeeding largely eliminates the negative correlation between infant mortality and subsequent birth spacing in Nepal.
Of particular interest are the studies that examine how breastfeeding beyond infancy affects mortality. Briend, Wojtyniak, and Rowland (1988) report that in Bangladesh children between the ages of 18 and 36 months who have been weaned have three times the mortality rate of those still
being breastfed; they attribute one third of the deaths in this age range to lack of breastfeeding. The World Health Organization (2000) estimates that in developing countries, mortality risk between ages one and two is twice as high if a child is not being breastfed.
Finally, other work focuses on health status and morbidity. Som, Pal, and Bharati (2007) report that in India shorter periods of breastfeeding are associated with a higher risk of stunting. Perera et al. (1999) find that breastfeeding reduces respiratory and diarrheal illness in Sri Lanka. Feachem and Koblinsky (1984) review over 30 studies conducted in 14 countries and find overwhelming evidence that breastfeeding reduces the risk of diarrheal disease.
7.2 Documented variation in child health across birth order and gender
Child health patterns with respect to birth order
Previous work on the relationship between birth order and child health appears to undermine our prediction that higher birth-order children, by virtue of nursing longer, enjoy better health out-comes. There are some proposed mechanisms that favor higher birth-order children (e.g., negative effects on first-borns from higher levels of intrauterine estrogen levels), but the medical literature primarily has identified mechanisms that favor lower birth-order children (Arad et al., 2001).
Many of the mechanisms cited by researchers relate to resource allocation. For example, Garg and Morduch (1998) find that in Ghana higher birth-order children experience more stunting and are more likely to be underweight than their older siblings, suggesting parents provide more calories during infancy to children born earlier. Behrman (1988) uses food consumption data to show that parents favor lower birth-order children in India. Our own results from Table 7 suggest that parents vaccinate lower birth-order children significantly more than later children. Finally, the mechanical negative correlation between birth order and family size may impart a health advantage to low-parity children. Thus, the favorable circumstances of low-parity children in terms of resource allocation and family size may swamp any breastfeeding effects.
Child health patterns with respect to gender
Several of the breastfeeding patterns we have documented coincide with previously established features of gender differentials in mortality in India. A distinctive feature of the missing women problem in India is that the sex ratio (the ratio of boys to girls) increases considerably in childhood; in China, by contrast, the problem is almost fully realized at or shortly after birth (Das Gupta, 2005). In India, girls have a forty percent higher mortality rate, relative to boys, between the ages of one and five but an equal mortality rate before age one (Acharya, 2004). This pattern coincides with our finding that most of the gender gap in breastfeeding does not arise until about twelve months after birth (Figure 3).
Furthermore, researchers have documented that in India excess female mortality is muted for first births (Das Gupta, 1987; Retherford and Roy, 2003). Similarly, we find relatively little gender difference in breastfeeding for first-born children.
7.3 Testing for a mortality-breastfeeding relationship in our sample
Empirical strategy
While the results from the existing literature are consistent with breastfeeding contributing to observed child mortality patterns, we can also directly test whether the breastfeeding patterns we found in the previous sections correspond to mortality differentials in the NFHS. The survey records mortality data for all children ever born to the mother, and we test whether the same gender, birth-order and ideal-family-size interactions that predicted breastfeeding duration in the previous sections also predict infant survival.
We use results from the previous sections as well as past research to construct tests of the mortality-breastfeeding hypothesis. The gender differences in the cumulative-distribution func-tion of breastfeeding durafunc-tion presented in Figure 3 indicate the age range where breastfeeding differences (and thus any related mortality differences) should be largest.20 The gender gap in breastfeeding is concentrated between the ages of 12 and 36 months, and there is no apparent gender gap during the first few months of the child’s life. Hence, we test for breastfeeding effects on the probability a child dies between 12 and 36 months after birth and use immediate post-natal mortality between zero and three months as a placebo test.
Furthermore, as the medical literature stresses breastfeeding’s benefits in the presence of un-sanitary water, we test whether the mortality patterns are most pronounced in households without piped water. Not only does this comparison shed light on the mechanism behind any health effects, it also allows us to separate the effects of breastfeeding from potentially confounding variables. For example, one important confounding factor is family size. The fertility-stopping patterns we observe in Table 5 imply that on average girls have more younger siblings, and evidence suggests that having a larger family size could lead to higher mortality rates for girls (Yamaguchi, 1989; Clark, 2000; Jensen, 2003; Rosenblum, 2008). We can make progress in separating the effects of family size and breastfeeding by comparing households with and without piped water, since family size should affect children in both types of households, but breastfeeding should primarily benefit children without piped water.
20There may also be long-term effects on the child’s health which lead to mortality at later ages; in essence we test only for contemporaneous
